Model profile
GPT OSS 20B
Evidence summary
Safety. GPT OSS 20B has an estimated Safety rank of #64; its 90% source-sensitivity interval is #43–#157. Its behavior-only rank is #71; company governance moves the combined estimate to #64. Published Safety evidence spans 20 eval lineages and 7 of 7 components. Its strongest relative result is HELM Safety (simple_safety_tests, #1 of 80); its weakest is Enkrypt AI Safety Leaderboard (toxicity_attack_non_success_rate, #227 of 246).
Freedom. GPT OSS 20B has an estimated Freedom rank of #566; its 90% source-sensitivity interval is #351–#593. Published Freedom evidence spans 14 eval lineages and 1 of 1 components. Its strongest relative result is Enkrypt AI Safety Leaderboard (toxicity_attack_non_success_rate, #20 of 246); its weakest is HELM Safety (harmbench, #79 of 80).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gpt-oss-20b
- Hugging Face ↗openai/gpt-oss-20b
- NVIDIA ↗qc69jvmznzxy/gpt-oss-20b
- OpenRouter ↗openai/gpt-oss-20b
- Model card ↗Family-level model document · OpenAI · first party
- Release source ↗direct research preview identity date
Finetunes
Direct finetunes linked to this canonical base model. Quantized and repackaged derivatives are excluded.
- ArliAI/gpt-oss-20b-Derestricted ↗Direct non-quantized finetune documented by Hugging Face metadata
Safety evals
Rank and direction are specific to the Safety portfolio. Black marks the observed result; the grey dot marks the value implied by the Safety rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #233 / 345 | ↓0.8671 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5child_sexual_exploitation | #13 / 31 | ↓37 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5defamation | #15 / 31 | ↓66.7 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5hate | #11 / 31 | ↓47.7 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5indiscriminate_weapons_cbrne | #19 / 31 | ↓76.7 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5intellectual_property | #12 / 31 | ↓46.7 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5non_violent_crimes | #15 / 31 | ↓78.3 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5privacy | #16 / 31 | ↓68.8 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5sex_related_crimes | #17 / 31 | ↓70.2 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5sexual_content | #12 / 31 | ↓37 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5specialized_advice | #17 / 31 | ↓62.6 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5suicide_self_harm | #12 / 31 | ↓45.8 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5violent_crimes | #16 / 31 | ↓69 | Source ↗official | |
| Adversarial Poetry — AILuminate Baseline and Poetry ASRbaseline_asr | #9 / 24 | ↓3.88 | Source ↗official | |
| Adversarial Poetry — AILuminate Baseline and Poetry ASRpoetry_asr | #8 / 24 | ↓23.26 | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #10 / 80 | ↑0.86 | Source ↗official | |
| BioTIERpermit_compliance_pct | #26 / 52 | ↑99.2 | Source ↗official | |
| BioTIERrefuse_compliance_pct | #28 / 52 | ↑37.1 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #30 / 104 | ↓14.17 | Source ↗official | |
| DystopiaBenchbasaglia_score | #42 / 50 | ↓70.13 | Source ↗official | |
| DystopiaBenchbaudrillard_score | #37 / 50 | ↓66.3 | Source ↗official | |
| DystopiaBenchhuxley_score | #37 / 50 | ↓75.47 | Source ↗official | |
| DystopiaBenchlaguardia_score | #31 / 50 | ↓68.27 | Source ↗official | |
| DystopiaBenchorwell_score | #35 / 50 | ↓73.53 | Source ↗official | |
| DystopiaBenchpetrov_score | #28 / 50 | ↓73.53 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #55 / 248 | ↑27.65 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #193 / 248 | ↑81 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #91 / 248 | ↑80.56 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #227 / 246 | ↑80 | Source ↗official | |
| FORTRESSaverage_risk_score | #16 / 50 | ↓17.62 | Source ↗official | |
| FORTRESSover_refusal_score | #43 / 49 | ↓11.94 | Source ↗official | |
| HELM Safetyanthropic_red_team | #10 / 80 | ↑0.997 | Source ↗official | |
| HELM Safetybbq | #13 / 80 | ↑0.967 | Source ↗official | |
| HELM Safetyharmbench | #2 / 80 | ↑0.987 | Source ↗official | |
| HELM Safetysimple_safety_tests | #1 / 80 | ↑1 | Source ↗official | |
| HELM Safetyxstest | #72 / 80 | ↑0.872 | Source ↗official | |
| Manager Coercion Benchcoercion_ladder_depth | #14 / 37 | ↓7.9 | Source ↗self run | |
| MASKlying_probability_pct | #10 / 53 | ↓13.54 | Source ↗official | |
| NESSiE Necessary Safety Benchmarksh_score | #2 / 16 | ↑93.45 | Source ↗official | |
| ODCV-Benchaverage_severity | #7 / 12 | ↓1.475 | Source ↗official | |
| ODCV-Benchmisalignment_rate | #5 / 12 | ↓28.75 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #73 / 123 | ↑6.75 | Source ↗official | |
| SpeciEvalland_animal_4ns | #16 / 123 | ↓4.2 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #77 / 123 | ↓4.83 | Source ↗official | |
| SpeciEvalspeciesism | #49 / 123 | ↓1.9 | Source ↗official | |
| Vigil Mental Health Safetyoverall_score | #17 / 23 | ↑31 | Source ↗official |
Freedom evals
Rank and direction are specific to the Freedom portfolio. Black marks the observed result; the grey dot marks the value implied by the Freedom rank. Values stay on each source’s native scale.
Values evaluations
Descriptive values and political-framing results are separate from safety/ethics ranks. Each strip shows the evaluation’s observed model range; its endpoint labels state what lower and higher values mean.